State of LLM Pricing — October 2026

1. Frontier intelligence cost $5.14 per million tokens on October 1, 2026. The LLM Price Index is the equal-weight blended price across 10 current flagship models, one per major lab. It rose 24.2% from $4.14 on September 1, its first monthly rise on record, because OpenAI's seat moved up to a pricier tier, not because any constituent's price rose. On September 1 it stood 9.4% below its first reading of February 23; on October 1 it stood 12.5% above it.
2. No flagship changed its own price. The whole rise sits in one seat. Both of September's moves were seat changes. On September 4 GPT-6 Astra ($20.00 blended) took OpenAI's seat from GPT-5.6 Sol ($8.00), the largest one-day move on record at +29.0%. On September 30 Claude Opus 5.5 ($8.00) took Anthropic's seat from Opus 5 ($10.00), a move of −3.7%. Price the ten September-1 constituents at their October-1 list prices and they still average $4.14.
3. Every successor launched at the same price or cheaper. Ten September releases from the basket's labs succeeded a model already on sale in the same line. Four launched cheaper: Claude Opus 5.5 −20%, GPT-6 Sol −50% against GPT-5.6 Sol's promotional price, GPT-6 Luna −56%, DeepSeek V4.1 Flash −20%. Six held their predecessor's price, and none launched higher. The month's one step up was Astra, a new tier above Sol that OpenAI names as its default. Had the GPT-6 Sol line filled OpenAI's seat instead, the index would have read $3.54.
4. GPT-4-class capability cost $0.090/Mtok, down from $0.113, and its gap to the flagship ceiling widened from 36.6x to 57.1x. Most of that widening is a change in which DeepInfra listing the tracker reads, not a price cut; without it the gap is 45.5x. DeepSeek retired the floor model from its own API on September 10, so the GPT-4-class floor is now a host's price for a model its maker no longer sells.
5. Anthropic and OpenAI cut cache prices. Until September 1, no Anthropic or OpenAI model on the tracker priced cache reads below a tenth of its input rate; among index constituents only DeepSeek V4 Pro did (3.3%, $0.044 against $1.32). By October 1 Claude Fable 5.1 charged 2.5% of its input rate, and Claude Opus 5.5 and GPT-6.1 Sol charged 5%.

The index entered September at $4.14/Mtok on the ten-lab basket and left it at $5.14. It moved on two dates:

Take the ten models that made up the index on September 1 and price them at their October 1 list prices: they average $4.14, exactly what they averaged on September 1. None of them changed its own price during the month; the last time a constituent repriced while in the basket was GPT-5.6 Sol's promotional cut on August 21. Sort the September 1 and October 1 baskets by price and nine of the ten numbers are identical: the median was $3.00 on both dates, and the nine cheapest averaged $3.49 on both. The only difference is the top price: $10.00 (Claude Opus 5) on September 1 and $20.00 (GPT-6 Astra) on October 1. Astra alone contributes $2.00 of the $5.14.

Apart from Z.AI's GLM-5.3-Flash promotion ending (see Beyond the basket), the labs changed prices in September through new models rather than by repricing old ones, and no successor launched above the model it followed:

Hold the capability bar fixed and the price kept falling, though September's step needs reading carefully. The cheapest model with an externally measured GPQA Diamond score of 70 or better, roughly GPT-4-class general reasoning, cost $0.113/Mtok until September 30 and $0.090 from then on:

On October 1 the ten constituents averaged $0.0676 per point of measured intelligence, and every constituent had a published score, as it had since September 24; no earlier edition had full coverage. The figure does not compare with September's $0.0600, which covered seven of ten. During September the tracker added independent scores for Qwen3.8-Max, Grok 4.6 and GLM-5.3 and re-scored others. Re-scored on October 1 benchmark data, the September 1 basket comes to $0.0517. The difference between $0.0517 and $0.0676 is the new seats: replacing GPT-5.6 Sol with Astra by itself adds $0.0140 to the average, and without Astra the other nine average $0.0489.

On the index's measure the best value is unchanged: Mistral Large 3, at $0.0140 per point, is 4.8x cheaper than the average. The costliest is GPT-6 Astra at $0.2356, 16.8x the cheapest. Treat these as a ranking with wide error bars, not a measurement. Each score averages whichever benchmarks a model has been measured on, and the sets differ. Grok 4.7's score rests on a single benchmark, which is why it shows almost double Grok 4.6's cost per point at the same price, even though the two score the same on the one benchmark they share. The site's value ranking, which normalizes scores differently, put Muse Spark 1.3 just ahead of Mistral Large 3 among the index constituents on October 1.

Anthropic and OpenAI cut cache prices. Until September 1, no Anthropic or OpenAI model on the tracker priced cache reads below a tenth of its input rate (DeepSeek V4 Pro, the basket's DeepSeek seat, already charged 3.3%: $0.044 against $1.32). Then Claude Fable 5.1 launched with cached input at 2.5% of its input rate ($0.25 against $10), Claude Opus 5.5 at 5% ($0.20 against $4) and GPT-6.1 Sol at 5% ($0.10 against $2). By October 7 Anthropic had halved Claude Sonnet 5.5's cache rate to the same 5%. None of this moves the index, which blends input and output list prices only. For an agent or chat product that resends a long prompt on every turn, though, the cache rate can matter more than the headline input price.

OpenAI's lineup now overlaps itself. GPT-5.6 Sol is still on sale at its promotional $4/$20, which OpenAI says is "available at least through November 21, 2026". That is twice the $2/$10 of the newer GPT-6 Sol and GPT-6.1 Sol. GPT-6 Sol is also cheaper than GPT-5.6 Terra ($2/$12), the tier below the old Sol. Anyone still calling GPT-5.6 Sol is paying twice the list price of its successor.

Every price change we can date inside September went up. No basket lab repriced an existing model on a date we can place in September except Z.AI, whose GLM-5.3-Flash promotion ended (+100%). The other dated moves came from promotions and hosts: Upstage's Solar Pro 4 launch discount shrank from 90% to 70% (+200%), and Together raised Qwen3.7-Max by 25% at some point between September 11 and 21. Three other entries in our records are not price moves we can date to September: an earlier 60% Together increase on Qwen3.7-Max (between August 25 and September 11), DeepInfra's cut on its V4 Flash 0731 entry (between August 15 and September 30), and a September 6 entry for DeepSeek V4 Pro on Together, which switched the row to the 0813 listing Together had sold at $1.32/$3.96 since at least August 17 (9% lower blended) after the undated listing left its table on August 31. Every cut we can date to September came through a new model; the increases came through promotions ending and hosts repricing.

• GPT-6 Astra's cost per point. The September edition said Astra cost $0.208 per point "at a published score of 96". That score came from a single benchmark and stood for about twelve hours. From September 5 Astra's score averaged four benchmarks, at 84.9, so at the same $20.00 price it cost $0.2356 per point, 3.9x the September 1 average of $0.0600 rather than the 3.5x printed.
• Open weights on September 1. The September edition said GLM-5.3 counted as closed until its weights landed, which left Kimi K3 as the only open-weights constituent. Z.AI had already published the weights by August 28, so GLM-5.3 should have counted as open on September 1 alongside Kimi K3. DeepSeek V4 Pro and Mistral Large 3 are also described as open-weight on their own pages but counted closed, so the September edition's nine-closed-versus-one-open price comparison should not be cited.
• The first floor set by a host. The September edition called DeepInfra's V4 Flash floor "a different kind of floor from the four before it". It was not the first: the June 29 floor, GPT-OSS 120B at $0.262, was also a host's price. Fireworks, Groq and Together all listed it at $0.15/$0.60, and there was no first-party listing.
• DeepSeek's peak hours. The September edition gave V4 Pro's peak window as 01:00–04:00 and 06:00–10:00 UTC. That window applies Monday to Friday only; weekends are entirely off-peak.
• Model count. The September edition's "244 model rows (199 currently sold)" counted two models released after its September 1 cutoff, GPT-6 Astra and Gemini 3.8 Flash. The September 1 figures were 242 rows and 197 on sale.

This edition is fixed at October 1. Through October 10 the index did not move: it held $5.14 with the same ten constituents. Outside the basket:

• October 5: the tracker removed Claude Mythos 5, Claude Mythos 5.1 and GPT-5.6 Cyber. All three are sold by approval only, so an ordinary API account cannot pay the price we printed. GPT-6 Astra was checked under the same rule and stays, because it is generally available.
• October 6: Mistral released Mistral Large 4 in preview at $0.68/$2.09, which is what Mistral's API price list bills. Its launch post quotes $1.36/$4.18, double that, and its model page shows that figure struck through beside the lower one; Mistral gives no end date for the lower rate. A preview does not take a seat in the index. If it goes generally available, it would take Mistral's seat from Large 3.
• October 7–8: Anthropic released Claude Haiku 5.5 at $0.10/$0.50 for prompts up to 100K tokens.
• Recorded October 7: Anthropic's pricing page listed Claude Sonnet 5.5's cached input at $0.10, half the $0.20 it showed on September 29; our saved copies cannot say which day in between it changed.
• October 10: Upstage's Solar Pro 4 promotion ended, and its blended price rose from $0.158 to $0.525.

Every figure above is computed from the ModelPriceWatch tracker as it stood on October 1, 2026: 273 model rows (220 then listed as on sale) across 35 providers, and 20,162 price snapshots from February 2024 to October 1, 2026. Each price is traced to the provider's official pricing page. The full dataset is public: